TL;DR — As of our last verification (2026-09-13), Claude Sonnet 5 bills $2.00 input / $10.00 output per million tokens — and that rate is now permanent. It launched as an introductory price due to rise to $3/$15 on September 1, 2026; Anthropic cancelled the increase, so the $2/$10 is simply the price. Sonnet 5 undercuts GPT-5.4 ($2.50/$15) on both sides and sits exactly 2× Haiku 4.5 and 2.5× below Opus 5. Note that most comparisons still online describe this as an expiring promo — check the as-of date on anything you read, including ours (pricing tracker).
The rate card (per 1M tokens, USD, verified 2026-08-26)
| Model | Input | Output | Note |
|---|---|---|---|
| Claude Sonnet 5 | $2.00 | $10.00 | Permanent — planned Sept 1, 2026 rise to $3/$15 cancelled |
| Claude Opus 5 | $5.00 | $25.00 | Frontier workhorse |
| Claude Haiku 4.5 | $1.00 | $5.00 | Value tier |
| GPT-5.4 | $2.50 | $15.00 | Closest OpenAI tier |
| Gemini 3 Pro | $1.25 | $5.00 | Value-frontier competitor |
Anthropic's discounts stack on top: ~90% off cached input on stable prompt prefixes and −50% via the batch API. A cached, batched Sonnet 5 call can land near Haiku-class effective pricing — which is why cache discipline matters more than the sticker price for agent workloads.
What a real month costs
At a standard workload of 10M input + 2M output tokens/month:
- Sonnet 5: $40.00
- Haiku 4.5: $20.00 · GPT-5.4: $55.00 · Gemini 3 Pro: $22.50 · Opus 5: $100.00
Run your own input/output mix through the LLM cost calculator — output share moves the answer more than most teams expect.
Where Sonnet 5 fits in a routing ladder
- Default reasoning tier for everyday synthesis, structured writing, multi-constraint answers — the rung between Haiku (bulk) and Opus (escalation)
- Head-to-heads: vs GPT-5.4 (Sonnet wins price outright) · vs Gemini 3 Pro (Gemini holds the price edge; quality on your evals decides) · vs Haiku 4.5 (the clean 2× step) · vs Opus 5 (the escalation pair)
- The cancelled increase is itself the routing lesson: teams that hard-coded a Sept 1 re-evaluation are now re-tuning against an event that never happened, while teams reading live pricing simply kept the cheaper rate. Price calendars are guesses; a router that reads current rates does not need them.
Routing decision card
- Route here when: the task needs real reasoning but not frontier-grade stakes — the bulk of "thinking" traffic in most agent stacks
- Escalate up (Opus 5) when: correctness is the budget — code that ships, multi-step agent chains, analysis someone acts on
- Route down (Haiku 4.5) when: classification, extraction, short chat turns, latency-sensitive calls
- Watch: Sept 1, 2026 — re-check the tracker before trusting any Sonnet 5 comparison published in August
ClawRouters keeps Sonnet 5 in its routing pool and applies the current rate per request, so the promo — and whatever follows it — is priced in automatically. All 24 Anthropic/OpenAI/Google/DeepSeek/GLM/Kimi pairs are indexed in the model comparison index.
Pricing sourced from Anthropic's official rate card as of 2026-08-26. Promotional terms are Anthropic's; verify the live rate at anthropic.com/pricing before committing volume.